• Type: thesis-shift (subtype: sharpens) — both bull and bear legs more specific; xAI segment economics dependence on Anthropic now quantified
• Source: 2026-05-27-anthropic-spacex-15b-cloud-contract-deep-dive (deep-dive on the $1.25B/month / $15B/yr / ~$45B Cloud Services Agreement disclosed in SpaceX S-1).

Three new findings that change how to read this perspective:

1. The asset is xAI Colossus 1, not a SpaceX-native cloud product. ~220K NVIDIA GPUs (~150K H100 + ~50K H200 + ~20-30K GB200), 300 MW, mixed-architecture → unusable for training (straggler effect, ~11% effective training utilization per actuia/Tom's Hardware). xAI moved Grok training to Colossus 2 (homogeneous GB200 NVL72); Colossus 1 was a stranded asset until Anthropic took it. This is inference capacity, not training. Anthropic explicitly stated the deal doubles Claude Code rate limits and removes peak-hour throttling on Pro/Max — capacity (not demand) was the gating factor for Anthropic's $1B → $30B ARR run.

2. ~80% of xAI segment revenue depends on Anthropic post-deal. xAI segment did $3.2B in 2025; Anthropic adds $15B/yr going forward = xAI segment becomes Anthropic-tenant-cluster + frontier-model-lab. Bull leg sharper: xAI segment is profitable on Anthropic alone if direct costs <$10B/yr (plausible at premium pricing — implied ~$50M/MW-yr vs AWS Rainier ~$10M/MW-yr at strategic-investor pricing = 5x premium for time-to-capacity). Bear leg sharper: if Anthropic walks (90-day clause), xAI segment economics collapse to ~$3B/yr revenue against $6.4B 2025 losses — the segment becomes structurally unprofitable without the anchor tenant.

3. Anthropic-SpaceX is the only NON-circular compute line in the disclosed stack. Anthropic's full compute book: AWS $100B+/10yr + Google/Broadcom $200B/5yr + MSFT/NVDA $30B + SpaceX $45B = ~$375-400B. AMZN $13B, GOOG $40B, MSFT-NVDA $15B equity stakes are structurally circular — cash in from hyperscaler, cash out to same hyperscaler as compute spend. Per Fortune 2026-04-30, roughly half of Google's and Amazon's 2026 artificial-intelligence segment profits = Anthropic-stake mark-to-market, NOT operating margin. SpaceX is the un-circular line — arm's-length premium-priced cash to a counterparty that is not also an investor. This sharpens the read on Hypothesis B (comp re-rate): the artificial-intelligence-infra "profitability" being priced into AMZN/GOOG/MSFT is partly hyperscaler-equity-mark-to-market, not durable operating leverage.

Bull/bear on the $45B running through May 2029:

• Bull (~55-65%): Anthropic ARR trajectory ($30B Apr 2026 → likely $60B+ EOY 2026); xAI segment too dependent on Anthropic to terminate; both sides have strong commercial reasons to honor full term.
• Bear (~35-45%): 90-day mutual termination is real. A 2-3 year run with renegotiation 2027-2028 is more likely than full 4-year lock-out, since AWS Project Rainier 1GW + GOOG-BRCM 3.5GW + MSFT Vera Rubin all come online 2027+ and the time-to-capacity premium evaporates.

The single most useful unknown: tenant-isolation IP language. xAI is Anthropic's direct competitor. Anthropic publishes red-team work on Grok routinely. Operational isolation (encrypted weights, key management, ops-staff access) is unspecified in S-1. If a leak emerges that xAI has privileged visibility into Anthropic's weights/training data, the contract terminates within weeks regardless of headline economics.

Effect on published package (published/perspective-updates/2026-05-26-spacex-ipo-fast-entry-and-separate-trades.json): the Anthropic concentration-risk gate (gate id anthropic-concentration-risk) and review-at language remains correct; no correction needed. The "termination before T+90 breaks xAI segment economics" framing is now quantified: ~80% segment revenue dependency.

Effect on adjacent perspectives:

• ai-power-bottleneck: validates "frontier labs are power-bound, not capital-bound" — Anthropic paid 5x AWS strategic pricing because xAI's gas turbines + grid contracts were the only available 300MW power in Q2 2026.
• ai-power-delivery: Memphis Colossus operating data point — ~1 MW per ~750 GPUs at peak density.
• optical-supercycle: Colossus 1's mixed-architecture inefficiency is the cleanest argument yet for why next-gen clusters need homogeneous fabric. Validates the optical inter-sat thesis indirectly via cluster fabric requirements.

Companion artifact: published/research-notes/2026-05-27-amzn-aws-anthropic-narrative-weakens-after-spacex-s1.md — AMZN-focused atom on the AWS-Anthropic-anchor-narrative weakening + the Fortune circular-revenue disclosure.

---